Re: [RFC PATCH v6 6/9] media: tegra: Add Tegra210 Video input driver

[Sowjanya Komatineni]

On 4/8/20 12:38 PM, Sowjanya Komatineni wrote:
> On 4/8/20 11:58 AM, Sowjanya Komatineni wrote:
> > On 4/8/20 10:45 AM, Sowjanya Komatineni wrote:
> > > On 4/8/20 7:21 AM, Dmitry Osipenko wrote:
> > > > External email: Use caution opening links or attachments
> > > >
> > > >
> > > > 08.04.2020 03:00, Sowjanya Komatineni ÐÐÑÐÑ:
> > > > ...
> > > > > > > > > I suppose that taking a shot takes at least few milliseconds, 
> > > > > > > > > which
> > > > > > > > > should be unacceptable to waste.
> > > > > > > > As long as buffers are in queue we have to keep processing each
> > > > > > > > buffer and between buffers obviously we have to wait for previous
> > > > > > > > frames to finish and this why we have separate thread for frame
> > > > > > > > finish where we can have next buffer capture ready and issue while
> > > > > > > > previous frame memory write happens
> > > > > > Also we specified numbers buffers as 3 to vb2 queue. So this is rare
> > > > > > case but to prevent issuing more than 2 at a time as VI HW is only
> > > > > > double buffered and syncpt fifo max depth is 2 added this to be 
> > > > > > safer.
> > > > > To be more clear, when more buffers are enqueued from userspace 
> > > > > always
> > > > > capture list will be full and thread will be busy in capture till 
> > > > > either
> > > > > error or stop stream request happens.
> > > > If kthreads take more than 1% of CPU time during capture (video) with
> > > > more than 2 buffers in queue, then it's not good and I think you 
> > > > should
> > > > do something about it. If kthreads stay at ~0%, then it should be okay
> > > > as-is.
> > >
> > > VI outstanding requests max can only be 2Â as syncpt fifo depth is 
> > > 2Â and waiting to issue next capture when already 2 captures are 
> > > inflight happens only during beginning of streaming where buffers 
> > > allocated go thru capture for first time after queuing.
> > >
> > > same buffers are returned to userspace after capture and same 
> > > allocated buffers will be queued back for subsequent captures.
> > >
> > > So this case of holding to issue single shot when already single 
> > > shot is issue for 2 frames simultaneous happens only during 
> > > beginning of start stream and also we set num_buffers to allocate 
> > > for queue as 3 although 2 is good enough where we will not hit this 
> > > case even during streaming start with 2 buffers
> > As 2 buffers are good enough to be clear will update in v7 to use 2 
> > buffers so we don't need to check for more than 2 outstanding buffers.
>
> correction: With 3 buffers, as soon as buffer is available capture 
> starts. So right most times I see it waiting for few ms before 3rd 
> capture to get through.
>
> As only 2 frames single shot can be issued in sequence (inflight 
> requests), instead of waiting for 1 of the request to finish, we can 
> use 2 buffers and avoid waiting as 2 buffers are good enough. Will 
> change this in v7.
Tested with 3 buffers and by checking outstanding buffers in process by 
VI hw and holding to start capture till one outstanding buffer in 
process by HW.
Also tested with 2 buffers without checking for outstanding buffers.

In both cases, I see same %CPU for the kthreads and is < 1%